Workshop on software strategies for sustainable physical modelling

Themes

• Programming Models

  For decades, Fortran has been the foundation of numerical weather prediction and climate modelling, underpinning systems such as IFS and ARPEGE. However, the emergence of GPUs, heterogeneous HPC systems, and domain-specific accelerators is prompting the community to reconsider how scientific software should be developed and optimised for future platforms.

  Increasingly, the discussion extends beyond programming languages towards higher-level programming paradigms that aim to separate scientific algorithms from hardware-specific implementation. Domain-specific approaches, such as the GT4Py4 and PSyclone DSLs (domain-specific languages), performance-portability libraries like Kokkos or  source-to-source frameworks like Loki and Fxtran, seek to improve portability, maintainability, and productivity while retaining access to modern HPC architectures. At the same time, frameworks such as PyTorch and JAX have demonstrated the power of expressing complex numerical algorithms through high-level abstractions that can be automatically differentiated and mapped onto diverse computing platforms, contributing significantly to the rapid pace of innovation in machine learning and scientific computing.

  This evolution raises important questions for the Earth-system modelling community. What level of abstraction is desirable to balance performance, portability, and long-term sustainability? Can common frameworks and programming models reduce the cost of adapting scientific codes to new architectures? How should large modelling systems be structured to take advantage of these emerging technologies?

  At the same time, ongoing modularisation efforts may reduce the need for a single technology stack across an entire forecasting system. Different components may increasingly be implemented using the most appropriate programming models, frameworks, or languages, provided they can be integrated through stable interfaces and common software engineering practices. The use of large language models and agentic code refactoring might also reduce constraints that previously prohibited the adaptation of novel programming paradigms increasing the need for robust interfaces and modular testing infrastructure.

  We invite contributions exploring programming abstractions, performance-portable development approaches, and experiences adapting scientific applications to emerging computing architectures. Relevant topics include domain-specific languages, source-to-source transformation, portability frameworks, heterogeneous programming models, software productivity, and maintainability.

• Agentic Coding and AI-assisted Development
  The rapid emergence of large language models and agentic software development tools is transforming how scientific software is written, maintained, documented, and tested. Capabilities that only recently required significant developer effort—including code generation, refactoring, documentation, testing, translation between programming languages, and performance analysis—can now be partially automated. While these technologies remain at an early stage, they have the potential to substantially influence the development of future Earth-system modelling systems.

  For scientific software communities, the implications extend beyond productivity gains. AI-assisted tools may lower barriers to entry, support knowledge transfer, accelerate modernisation of legacy codes, and facilitate the adoption of new programming models and architectures. At the same time, they raise important questions regarding software quality, reproducibility, governance, validation, attribution, and the long-term maintenance of complex scientific applications.

  Particularly interesting are emerging agentic workflows in which AI systems are able to perform multi-step development tasks such as implementing new functionality, generating tests, reviewing code changes, or adapting software to new computing architectures. Such approaches may fundamentally alter how software is developed and maintained, potentially enabling smaller research and operational teams to manage increasingly complex software ecosystems.

  We invite contributions exploring the opportunities and challenges of AI-assisted software development in scientific computing. Relevant topics include code modernisation, automated testing and verification, software maintenance, performance portability, developer productivity, governance of AI-generated contributions, and practical experiences deploying these technologies in research or operational environments.

• Composition and sharing of Earth System Models

  Earth-system modelling is entering a period of rapid transformation. Forecasting systems are becoming increasingly diverse, spanning traditional numerical weather prediction models, Earth-system components, data assimilation systems, and emerging data-driven approaches. At the same time, the scientific and technical expertise required to develop and maintain these systems is becoming more specialised and distributed across organisations.

  A key question for the coming decade is how the community can best organise the development and sharing of software infrastructure. The traditional model of large monolithic code bases is increasingly being complemented by modular architectures built around reusable components, clearly defined interfaces, and shared software ecosystems. Such approaches can improve maintainability, portability, testing, and adaptation to new computing architectures, while also enabling broader collaboration between operational centres, academia, and industry. Recent community initiatives, including the collaborative development model adopted by the Anemoi framework, illustrate how shared infrastructure can accelerate innovation while avoiding unnecessary duplication of effort. [Anemoi]

  This raises important questions for the wider Earth-system modelling community. To what extent should common infrastructure, software libraries, workflows, and development practices be shared across organisations? How can independently developed components be combined into coherent forecasting and data assimilation systems while preserving scientific integrity, reproducibility, and operational robustness?

  The growing convergence between physical and data-driven methods provides an additional dimension to this discussion. Many of the challenges surrounding data management, workflow execution, training and inference infrastructure, verification, and operational deployment are common across both paradigms. Exploring opportunities for shared infrastructure may help foster greater interoperability between physical and machine-learned components, support hybrid modelling approaches, and create more sustainable software ecosystems for future forecasting systems.

   

• Workflow Orchestration and Operational Infrastructures
  

  Forecasting systems are increasingly expected to operate across a diverse ecosystem of computing environments, ranging from traditional HPC platforms and emerging accelerator-based systems to cloud infrastructures, edge computing environments, and hybrid deployments. At the same time, the software workflows underpinning forecasting systems are becoming more complex, spanning data acquisition, model execution, data assimilation, machine learning training and inference, verification, post-processing, and product generation. As scientific workflows evolve, there is a growing need to reconsider how these activities are coordinated, deployed, and managed across heterogeneous computing resources.

  An important aspect of this discussion is the emergence of data-driven forecasting methods, which often exhibit very different workflow characteristics from traditional numerical weather prediction systems. Physical forecasting systems typically revolve around tightly coupled operational cycles combining data assimilation, forecast production, product generation, and verification. Data-driven methods, by contrast, introduce additional workflows associated with large-scale data preparation, model training, re-training, experimentation, validation, and inference. These activities may have different computational requirements, timescales, software dependencies, and execution environments. As hybrid physical–ML systems become increasingly common, understanding how these distinct workflows can coexist and interact efficiently becomes a key challenge.

  Recent developments in both the Earth-system modelling and machine-learning communities have demonstrated the potential benefits of modern workflow technologies, including automated pipelines, cloud-native services, containerised applications, flexible orchestration frameworks, and reproducible deployment practices. However, many questions remain regarding the extent to which such capabilities should be developed independently within organisations or shared across the wider community. What level of common infrastructure, tooling, and operational practice would maximise innovation while avoiding unnecessary duplication of effort? Which components should remain organisation-specific, and which could benefit from community-wide collaboration?

  We invite contributions exploring approaches to workflow orchestration, deployment, and operational integration across traditional and emerging forecasting paradigms. Topics of interest include execution across heterogeneous computing platforms, cloud and hybrid-cloud infrastructures, reproducibility and portability, workflow interoperability, automation and continuous delivery, operationalisation of machine-learning systems, and architectures that support both physical and data-driven modelling approaches. We are particularly interested in experiences that help clarify where greater sharing, coordination, or consolidation of workflow infrastructure could enable more sustainable and agile forecasting systems for the future.

   

• Hybrid modelling and the Future Data Assimilation System
  

  A central challenge is how best to support hybrid modelling that combines physical and statistical approaches, including emulation of physical model components and data-driven model augmentation approaches. From a technical perspective, integration of ML poses challenges due to domain decomposition strategies intrinsic to physical models, which often conflict with the global or non-local architectures typical of ML emulators. Overcoming these challenges would require carefully designed interfaces that reconcile the spatial partitioning of numerical models with ML requirements, ensuring consistent data exchange, scalability, and performance.

  Hybrid modelling is also interesting for data assimilation, which will remain a fundamental component of future forecasting systems, not only because it provides high-quality training data for ML models, but also due to its ability to produce high-quality initial conditions. This calls for a flexible framework that enables seamless integration of ML models within the data assimilation process. ML integration could extend beyond hybrid models to important applications such as modelling background and observation error covariance structures with ML algorithms, as well as emulating observation operators, tangent linear, and adjoint models. These advancements aim to enhance assimilation performance while reducing computational costs through more efficient variational assimilation.

  By embracing this hybrid modelling approach, the system should support the simultaneous optimisation of physical states and ML-driven components for the forward model and within assimilation cycles, fostering innovations that combine established numerical methods with modern data-driven techniques.